Everybody and their uncle is writing about DeepSeek this week, mostly about the stock market stuff, so I'm going to skip that part entirely (you've read it four times already, I promise you have) and just tell you what happened when I actually tried to run the thing myself, on my own hardware, in my own kitchen, at around 11pm on a Sunday because I couldn't sleep and my brain does this sometimes.
Quick background for anyone who's been offline: DeepSeek is the Chinese lab that dropped R1, a reasoning model that's apparently competitive with OpenAI's o1 on a bunch of benchmarks, and unlike o1 they actually published weights you can download. Not just the giant 671-billion-parameter monster (which nobody's running on a laptop, don't be silly), but a whole family of smaller "distilled" versions built on top of Qwen and Llama base models, ranging from 1.5B all the way up to 70B.
Ollama added support almost immediately, which honestly still amazes me a little. Open source moves fast when it wants to.
Getting it running
If you already have Ollama installed this is stupidly easy. If you don't, go grab it from ollama.com, it's a five minute install on Mac, Windows, or Linux and I'm not going to walk through that part because it's just a normal installer.
Once it's in, you just do:
ollama run deepseek-r1:8b
That pulls down the 8-billion-parameter distilled version, which landed at about 4.9GB on my machine. I'm on a 2023 M2 MacBook Pro with 16GB of RAM, which is not a beefy machine by any modern AI standard, and it still churned through the download in under ten minutes on my home connection.
First response took maybe 4 seconds to start streaming, which felt totally normal. What's genuinely different from other local models I've messed with (and I've tried a lot of them, Llama 3, Mistral, that whole zoo) is that R1 shows its reasoning before giving you the answer. It literally writes out a <think> block where it argues with itself, second-guesses stuff, backtracks. I asked it a fairly annoying logic puzzle about three people and their ages and watched it talk itself in a circle for a solid paragraph before landing on the right answer. It's a little unsettling to watch, honestly, like reading someone's internal monologue when they didn't know you were in the room.
I also grabbed the 14b version just to compare. That one's noticeably slower on my hardware, maybe 2-3x, and the fan on my laptop kicked on almost immediately and stayed on for the better part of twenty minutes. My cat left the room. That's not a joke, she got up off the arm of the couch and left, which either means the pitch of the fan bothered her or she's developed opinions about AI safety and neither would surprise me at this point.
Is it actually good
For an 8B model running fully offline, disturbingly good. I wouldn't trade it for GPT-4 on anything that needs broad world knowledge, it's smaller and it shows. But for step by step reasoning tasks, math word problems, that kind of thing, it punches way above its weight class. I threw a few coding questions at it too and the 14b version wrote a working (if slightly over-commented) Python script for parsing a CSV on the first try, no follow up needed.
The thing I keep coming back to, though, isn't even the quality. It's that this is free, runs on hardware I already own, and works with the wifi off. I unplugged my router halfway through testing just to prove it to myself and kept chatting with it. After years of every "AI breakthrough" being locked behind an API key and a monthly bill, having a genuinely capable reasoning model sitting in a folder on my own SSD feels like a small, weird kind of freedom. I didn't expect to feel anything about a file download, but here we are.
One complaint: the smaller 1.5b and 7b versions are noticeably worse at following instructions and will occasionally just ignore part of your prompt entirely, so don't judge the whole family by the tiny ones. Go at least 8b if your machine can handle it, 14b if you've got 32GB of RAM or better and don't mind the noise.
Anyway. My laptop is currently sitting on a cooling pad I bought two years ago and never used until tonight. Some things come full circle eventually.